If anyone at your firm has pasted a client's documents into an AI tool, asked one to summarize a deposition, or used one to read text out of a screenshot, the supervision question is already live. ABA Model Rule 5.3 does not mention artificial intelligence by name, but it governs the nonlawyer assistance your lawyers rely on, and ABA Formal Opinion 512, issued in July 2024, applies that framework directly to generative AI.
This guide is written for the people who carry the supervisory duty: managing partners, practice group leads and the attorneys who sign what paralegals and tools produce. It covers what the rule requires at the firm level and at the desk level, how to build the measures the rule calls for, what verification looks like task by task, what to ask vendors, and where responsibility lands when supervision fails. State versions of Rule 5.3 and state bar AI guidance vary, so treat the Model Rule as the baseline and check your own jurisdiction's text and opinions.
What does Rule 5.3 require when your firm uses AI?
Rule 5.3 creates three layers of responsibility for nonlawyer assistance, and each maps onto AI use. The duty is one of reasonable efforts, not perfection, but it is affirmative: someone has to build the system and someone has to run it.
- Firm-level measures, 5.3(a). Partners and lawyers with comparable managerial authority must make reasonable efforts to ensure the firm has measures in effect giving reasonable assurance that nonlawyers' conduct is compatible with the lawyers' professional obligations. For AI, that means policies, an approved-tool list, training and vendor diligence.
- Direct supervision, 5.3(b). A lawyer with direct supervisory authority over a nonlawyer must make reasonable efforts to ensure that person's conduct is compatible with the lawyer's obligations. This is the attorney who assigns a paralegal a task and decides whether an AI-assisted result is good enough to use.
- Responsibility for violations, 5.3(c). A lawyer is responsible for a nonlawyer's conduct that would violate the rules if done by a lawyer when the lawyer orders or ratifies it, or when a lawyer with managerial or supervisory authority knows of it at a time when its consequences can be avoided or mitigated and fails to take reasonable remedial action.
Rule 5.1 runs in parallel for lawyers. An associate who uses an AI tool carelessly is supervised under 5.1, so the firm-level measures you build should cover lawyers and staff alike.
Who counts as the "nonlawyer" when AI is involved?
In practice there are three: the staff member using the tool, the vendor that provides it, and the output itself, which arrives looking like finished work. Each needs a different control.
The rule's title was changed in 2012 from "Assistants" to "Assistance," and its comments describe services outside the firm, such as document management companies and internet-based storage of client information. Opinion 512 looks at AI vendors through that same lens. The practical result is a three-part map:
| Layer | Example | Control that fits |
|---|---|---|
| The person using AI | A paralegal summarizes a client's message export with a chatbot | Policy, training, approved tools, assignment-level instructions |
| The vendor | A cloud service that stores and processes uploaded client files | Due diligence, contract terms, access controls, periodic review |
| The output | A draft chronology, extracted text, a summary or a citation | Verification against sources before anyone relies on it |
What does ABA Formal Opinion 512 add?
Opinion 512 makes the supervisory expectations concrete: managerial lawyers should establish clear policies on permissible AI use, and supervisory lawyers should make reasonable efforts to ensure lawyers and nonlawyers follow them. It also expects training on how the tools work, their limitations, and secure handling of client data.
The opinion ties supervision to the other duties it discusses. Competence under Rule 1.1, including comment 8's call to understand the benefits and risks of relevant technology, means the supervisor has to understand a tool well enough to know where it fails. Confidentiality under Rule 1.6 decides which tools may receive client information at all; the opinion concludes that informed client consent is needed before information relating to a representation goes into a self-learning tool that could disclose it, and that boilerplate in an engagement letter does not suffice. Rule 1.5 limits what you can bill for AI-assisted work. Our walkthrough of ABA Formal Opinion 512 covers each section; the rest of this article focuses on building the supervision system itself.
How do you build "measures giving reasonable assurance"?
Treat Rule 5.3(a) as a project with a deliverable: a written AI program that people are trained on and that gets reviewed. A workable version for a small or midsize litigation practice looks like this.
- Inventory what is already in use. Ask every lawyer and staff member which AI features they use, including ones built into email, PDF, transcription and practice management tools. The tools nobody approved are the ones that matter most.
- Classify uses by risk. Sort tasks into permitted, permitted with verification, and prohibited. Grammar help on a non-confidential draft is a different risk from uploading a client's medical records.
- Approve tools, not just uses. Keep a short list of vetted tools for client data, the settings that must be on or off, and who may use each one.
- Write verification into the workflow. Define, per task type, what a person must check before output is used, and who signs off.
- Train and document it. Train lawyers and staff on the policy and on each approved tool's limits. Keep the materials and attendance records.
- Review on a schedule and on triggers. Revisit the program periodically and whenever a vendor changes its terms, a new tool is adopted, or new ethics guidance issues in your jurisdiction.
Sample wording: AI use policy clauses
1. Approved tools. Information relating to a client representation may be entered only into tools on the firm's approved list, configured as that list specifies. Public or consumer AI tools are not approved for client information.
2. Verification. AI output is a draft. Before it is used in client advice, a filing, an exhibit or correspondence, the assigned team member verifies it against the source materials, and the supervising attorney approves its use.
3. Consent and disclosure. The responsible attorney decides whether client consent or disclosure is required for a given use, consistent with firm guidance and applicable rules.
4. Reporting. Anyone who discovers an AI-related error, a confidentiality concern or an unapproved use reports it to [name or role] the same day.
Adapt the wording to your firm's structure and your state's rules. It is a starting point, not a finished policy.
What does verification look like in practice?
Verification means checking output against the source it claims to reflect, by a person qualified to spot the error, before anyone relies on it. The right check depends on the task.
| Task | Typical failure | Minimum check |
|---|---|---|
| Text extraction from screenshots or photos | Misread names, times and numbers; dropped lines | Compare extracted text to the image; correct and record changes |
| Summaries of messages or transcripts | Missing context, merged speakers, overstated tone | Spot-check against originals; confirm every fact used downstream |
| Chronologies | Wrong dates, invented sequencing | Trace each entry to a source document or message |
| Legal research and citations | Nonexistent or mischaracterized authority | Pull and read every authority cited |
For a deeper look at failure modes, see preventing AI hallucinations in legal work, and share the paralegal's guide to supervision, verification and ethics with your staff so desk-level habits match the firm policy.
What should you ask an AI vendor before it touches client data?
Ask the questions the ABA has long associated with outsourcing: who can see the data, what the vendor does with it, how it is secured, and what happens when something goes wrong. Get the answers in writing.
- Does the vendor use your inputs or files to train models, and can that be ruled out by contract, not just a settings toggle?
- Where is data stored, who at the vendor can access it, and is that access logged?
- Will the vendor notify you of a breach, and of a subpoena or other demand for your clients' data?
- Can you delete data on demand and at the end of the relationship, and receive confirmation?
- Can access be scoped by user, role and matter, and can credentials expire and be revoked?
- If an AI assistant connects to your systems, can you limit it to read-only access or to specific functions?
The last two items matter more as firms connect AI assistants to case data through APIs and the Model Context Protocol. Look for scoped tokens with expiry and revocation and integrations that respect existing user permissions, so an assistant can never reach more than the person who connected it.
What happens when supervision fails?
Under Rule 5.3(c), the supervising lawyer can become responsible for a nonlawyer's conduct by ordering it, by ratifying it, or by knowing about it in time to fix it and not acting. With AI, ratification is the quiet risk: signing a filing or approving an exhibit built on unverified output can make the error yours.
When an error surfaces, the remedial steps mirror those for a human mistake: correct the record promptly, assess whether the client must be informed under Rule 1.4, document what happened, and fix the control that failed, whether that is training, the approved-tool list or the verification step. If you are unsure how your state's version of these rules applies to a specific situation, consult ethics counsel licensed in your jurisdiction or your state bar's ethics hotline before deciding.
